Beyond Remediation: An Irreversibility Threshold for Preventive AI Governance

Most AI governance frameworks share one assumption. If something goes wrong, it can be fixed afterward with audits, patches, or liability. This paper argues that assumption fails in three situations. AI generated misinformation that reshapes beliefs faster than corrections spread. Synthetic defamation that destroys reputations before any court can act. Automation that erases local labor markets by doing exactly what it was built to do. In these cases remediation arrives too late. The damage is already distributed, already embedded in institutions, or already inside the next training corpus. This work introduces the Systemic Irreversibility Baseline, a practical test built on three variables: non restorability, relational depth, and distributional diffusion. The framework routes AI harms into four governance tracks based on whether restoration is realistically possible. Fixable failures remain within existing oversight systems. Localized irreversible harms trigger strict liability and stronger intervention. Systemic product side harms activate mandatory design safeguards. Labor displacement moves to economic policy rather than product safety regulation. For irreversible product harms, the paper proposes four concrete technical mandates: tamper evident human authorization logs, conformal prediction based uncertainty controls, interface friction that forces engagement with model confidence, and C2PA based provenance infrastructure. The goal is not to replace current governance frameworks such as the EU AI Act, NIST AI RMF, UK White Paper, or OECD incident taxonomy. The goal is to add the missing threshold that signals when repair is no longer possible. When harm cannot be undone, governance must move upstream. Prevention replaces remediation.